counterfeit-V2.5 2.5d tweak
Run Model | 使用GRAVITI Diffus 关于GRAVITI Diffus |
Model Name | counterfeit-V2.5 2.5d tweak |
Model Type | Checkpoint |
Base Model | SD 1.5 |
Version Name | counterfeit-V2.5 2.5d tweak |
File Name | counterfeitV2525d_counterfeitV2525d.ckpt |
SHA256 | d3da621b47 |
NSFW | False |
Trigger Words | |
Tags | anime, style |
Metadata | fp = fp16 size = full format = PickleTensor |
Size | 7.17 GB |
Version Description | |
Source | https://civitai.com/models/10443 |
Author | pzzmyc |
这不是counterfeit 2.5!
与传统融合不同,本模型通过调整U-NET来达到类似orangemix的效果,实际上orangemix也是这样做的。
使用方法:
首先也需要EasyNegative这个pt文件,与840000VAE
与原版类似,反面tag越少越好,但建议加入关于手指和脚趾,手部脚部的反面tag。
step 30左右,hiresfix是必须的,分辨率越高越好,denoising 0.5-0.6
与原版对比缺点:
1.cf特有的填充画面所用的箱子丢失,画面元素丰富程度不如原版,
2.可能不适合nsfw(未测试,原模型nsfw也不好)
3.漫画笔触因为画风原因丢失。
与原版对比优点:
1.更写实的2.5d风格,
2.相同的种子构图基本与原模型一致,保留了70%左右的模型特色,包括服装,瓶瓶罐罐,小物件,可以自行测试
3.更加浓厚的色彩,不再发灰,
4.调节了语义层,能更好理解输入的tag,(玄学)
5.由于画面偏向写实,手脚表现理应更好
//////////////////////////////////////////////////////////////////////////////////////
This is not counterfeit 2.5!(I’ve tried my best to translate)
Unlike traditional fusion, this model adjusts U-NET to achieve an orangemix-like effect, which in fact does the same.
How to use:
First of all, you also need EasyNegative pt file, with 840000VAE
Similar to the original counterfeit, fewer negative tags is better, but it is recommended to include negative tags about fingers and toes,
Step 30 or so, hiresfix is a must, higher resolution is better, denoising 0.5-0.6,
compared to the original conterfeit:
1.original Model features like boxes used to fill the picture are missing, and the richness of the picture elements is not as good as the original,
2. May not be suitable for NSFW (not tested, the original model NSFW is not good too)
3. Comic strokes missing due to drawing style.
Advantages compared to the original:
1. More realistic 2.5D style,
2. The same seed of picture composition is basically consistent with the original model, retaining about 70% of the model features, including clothing, bottles and jars, small objects, which can be tested by yourself,
3. More intense color, no more gray,
4. Adjust middle layer to better understand the input tag, (metaphysics)
5. Since the picture is biased towards realism, the performance of hands and feet should be better
此外:
1.没有联系过原作者,懒,如果原作者有任何不满请留言,我立马删除
2.他妈滴7gb传了十个小时,百度网盘都比这破站快